Signal Over Noise: The three math problems AI still needs to solve

Ulrike Hoffmann-Burchardi · UBS Global Wealth Management

AI's addressable market is large enough to eventually justify the infrastructure build-out, though timing is uncertain.

As AI's cognitive abilities increase, the total addressable market is large: about 50 trillion in knowledge-work salaries and 5 trillion in factory-worker salaries. That size is the basis for the view that AI revenues can eventually justify the trillions of infrastructure spending. The speaker says the speed of that revenue build is unresolved ('How quickly remains to be seen'), so the case rests on market size rather than on a timeline.

“With AI's increasing cognitive abilities, the total addressable market is large: 50,000,000,000,000 knowledge and 5,000,000,000,000 factory workers' salaries. How quickly remains to be seen.”at 0:23

Ulrike Hoffmann-Burchardi · UBS Global Wealth Management

Past tech disruptions claimed casualties among both incumbents and new entrants, and AI is likely to follow the pattern.

Prior technology cycles show that disruption produces losers on both sides, incumbents and new entrants. The speaker notes that major AI contenders, both AI natives and incumbents, have large infrastructure commitments, so the same risk applies. The argument rests on historical precedent rather than on specific AI company data.

“Prior [tech] cycles have shown that disruptions tend to claim casualties on both sides: incumbents and new entrants.”at 4:17

Ulrike Hoffmann-Burchardi · UBS Global Wealth Management

Data center permitting is a worsening bottleneck, but projects will shift to friendlier states and mostly complete.

Permitting remains a bottleneck, and the new political climate after the US midterms makes it harder. The bypass is other jurisdictions, so new permits could shift toward data-center-friendly states such as North Dakota and Indiana. The speaker's base case is that data centers already under construction or already permitted will be completed. He also argues providers are changing their playbook to support local communities, which may reduce opposition below what was feared.

“Permitting of data centers remains a bottleneck and more so with the new political realities post the U. S. Midterms.”at 1:15

Ulrike Hoffmann-Burchardi · UBS Global Wealth Management

Financing is a human-controlled bottleneck that tightens if AI revenues arrive slower, slowing the build-out further.

Financing is a critical bottleneck that people control, and it becomes more constraining if AI revenues grow more slowly than expected. Slower revenue would throttle the build-out, which would in turn slow the pace of AI revenue, creating a feedback loop. The speaker adds that AI hopes can drive upside while fears of a slower build-out can drive downside, and that market psychology will likely be decisive.

“Lastly, financing remains a critical human controlled bottleneck. It could become more constraining if AI revenues move slower than anticipated.”at 2:02

Ulrike Hoffmann-Burchardi · UBS Global Wealth Management

Power is the tightest physical bottleneck, setting the pace of the AI build-out through long gas turbine lead times.

The supply chain for AI infrastructure is only as fast as its tightest bottleneck, and the speaker says power likely sets that pace. As evidence, gas turbines currently face lead times of about five years, against a normal one to two years. That delay limits how fast new AI data-center capacity can be energised.

“On the physical side, the supply chain is only as fast as its tightest bottleneck. Here, power likely sets the pace. For example, gas turbines face current lead times of five years compared to one to two years normally.”at 0:50

Ulrike Hoffmann-Burchardi · UBS Global Wealth Management

Semiconductor manufacturing lead times are stretched, making chip production the second physical bottleneck after power.

Semiconductor manufacturing 'follows closely behind' power as a physical constraint on AI build-out. Current lead times are three to four years, against a usual two to three years. Those longer lead times slow how fast AI infrastructure can be built.

“Semiconductor manufacturing follows closely behind. Current lead times are three to four years compared to the usual two to three years.”at 1:01

Ulrike Hoffmann-Burchardi · UBS Global Wealth Management

High inference margins do not guarantee AI profits, since rising training costs can erode them; scale favours concentrated, profitable markets.

Open-source models have shown gross margins of 85% on some inference workloads, but inference profitability is not company profitability, because inference margins exclude much of the cost of training frontier models. Even an 85% inference margin will not guarantee overall profit if the cost of developing the next model keeps rising. The speaker's argument on scale is that revenue lets companies spread fixed training costs, which lowers per-unit cost and widens net margins. Markets with high economies of scale tend toward concentrated structures with stable, high profitability, so scale and market structure will likely determine the outcome.

“Even an 85% inference margin doesn't guarantee overall profitability if the cost of developing the next model keeps rising.”at 3:09

Ulrike Hoffmann-Burchardi · UBS Global Wealth Management

Cheaper open-source models cap prices for workloads that do not need frontier intelligence, complicating AI profit economics.

Open-source models are cheaper and slightly less capable than frontier models. The speaker argues they can keep prices in check for workloads that don't need frontier intelligence, which complicates the question of whether AI revenue turns into profit. This pressure on pricing is one reason the speaker expects frontier labs to move into applications.

“Cheaper open source models with slightly fewer capabilities complicate this equation. They can keep prices at bay for workloads that don't need frontier intelligence.”at 3:46

Ulrike Hoffmann-Burchardi · UBS Global Wealth Management

Frontier models will increasingly move into AI applications, where they can build data, domain knowledge and customer moats.

Because cheaper open-source models pressure prices on workloads that don't need frontier intelligence, the speaker believes frontier models will expand into applications. In those applications, frontier models can accumulate proprietary data, domain knowledge and customer relationships, which the speaker presents as the source of their durable advantage.

“This is why we believe frontier models will increasingly expand into AI applications where they can build data, domain knowledge and customer modes.”at 3:58